Incorporating sub-programs as knowledge in program synthesis by PushGP and adaptive replacement mutation

Program synthesis aims to build an intelligent agent that composes computer programs to solve problems. Genetic programming (GP) provides an evolutionary solution for the program synthesis task. A typical GP includes a random initialization, an unguided variation, and a fitness-guided selection to search for a solution program. However, several recent studies have shown the importance of using prior knowledge in different components of the GP. This study investigates the effectiveness of incorporating sub-programs as "prior knowledge" into the variation process of GP by Replacement Mutation. We further design an adaptive strategy that allows the automatic selection of the helpful sub-programs to the search process from an archive (including helpful and unhelpful ones). With handcrafted sub-program archives, we verify the effectiveness of the Adaptive Replacement Mutation method in success rate. We demonstrate the effectiveness of our approach with transferred archives on two composite problems.